List your company

Long-context Retrieval on Needle-in-a-Haystack

100Retrieval AccuracyQwen2.5-3B-Instruct (Teacher)
-4235077Mar 29, 2025Jun 8, 2025Aug 19, 2025Oct 29, 2025Jan 9, 2026Mar 21, 2026Jun 1, 2026

Evaluation results

MethodLinks
Qwen2.5-3B-Instruct (Teacher)Context length (tokens)=8,1922025.12100
Hybrid student (25% softmax, 75% GDN)Context length (tokens)=8,1922025.12100
Qwen2.5-3B-Instruct (Teacher)Context length (tokens)=16,3842025.12100
Qwen2.5-3B-Instruct (Teacher)Context length (tokens)=32,7682025.12100
Qwen2.5-3B-Instruct (Teacher)Context length (tokens)=65,5362025.12100
USMoEContext Length=16.5K, Doc Depth=50%, Base Model=Qwen3-30B-A3B-Instruct2025.03100
TransformerScale=50M2026.06100
SISAScale=50M, ds=642026.06100
TransformerScale=152M2026.06100
SISAScale=152M, ds=162026.06100
TransformerScale=369M, mb=42026.06100
SISAScale=369M, ds=32, mb=42026.06100
SISAScale=369M, ds=64, mb=4/82026.06100
SISAScale=369M, ds=128, mb=42026.06100
Mamba-3Scale=369M, mb=22026.06100
Hybrid student (25% softmax, 75% GDN)Context length (tokens)=16,3842025.1299.8
Hybrid student (25% softmax, 75% GDN)Context length (tokens)=32,7682025.1299.8
SISAScale=369M, ds=32, mb=22026.0699.5
Hybrid student (25% softmax, 75% GDN)Context length (tokens)=65,5362025.1299.4
Mamba-3Scale=152M2026.0699
Mamba-2Scale=369M, mb=42026.0696.5
Qwen2.5-3B-Instruct (Teacher)Context length (tokens)=131,0722025.1295.4
Mamba-3Scale=369M, mb=42026.0686.5
Mamba-2Scale=152M2026.0682.5
Hybrid student (25% softmax, 75% GDN)Context length (tokens)=131,0722025.1268.4
Mamba-3Scale=50M2026.0663.5
Original (TC)Context Length=16.5K, Doc Depth=50%, Base Model=Qwen3-30B-A3B-Instruct2025.0350
ECContext Length=16.5K, Doc Depth=50%, Base Model=Qwen3-30B-A3B-Instruct2025.0310
Mamba-2Scale=50M2026.060